Conversation is an effective way for people to naturally connect and share thoughts. The Web hosts a vast amount of conversations that reflect diverse user needs and preferences. My research seeks to use large language models (LLMs) to enhance user experience regarding online conversations. This dissertation explores three problems.

First, users reach out to live chat agents for help; can we predict their intent in advance? We propose the problem of predicting user intent from browsing history and address it through a two-stage approach. The first stage fine-tunes a pretrained Transformer for intent classification, and the second stage uses LLMs to generate fine-grained intents. Our two-stage approach yields significant performance gains compared to generating intents without the classification stage.

Second, can we develop a single-turn chatbot that provides recommendations based on free-form natural language queries? LLMs on their own are limited by lack of external knowledge and inherent biases. As such, we propose an approach that equips LLMs with various tools that grants access to external knowledge bases and specialized models. This approach results in more factual, relevant, and diverse recommendations.

Finally, we envision multi-turn chatbots for recommendation. Since it is expensive to test chatbots with real users, simulators are often used as substitutes. How realistic can LLMs be as simulators? To address this question, we introduce a protocol to measure the degree to which LLMs can accurately emulate human behavior in conversational recommendation. These tasks reveal deviations in language models from human behavior and guide us on reducing these deviations.